# Baidu releases ERNIE-5.5-Flash-Base, a 280B MoE base model trained on 33.5T tokens

> Baidu released ERNIE-5.5-Flash-Base, a base model with 280B total and 14B active parameters, and reports gains over DeepSeek-V4-Flash-Base.

- Status: reported (One outlet, not independently verified yet.)
- Topic: Models
- Published: 2026-11-07T03:06:06+00:00
- Updated: 2026-11-07T03:06:06+00:00
- Web page: https://ai-news.ikangai.com/story/baidu-releases-ernie-5-5-flash-base-a-280b-moe-base-model-trained-on-33-5t-token/
- Seen in: Latent Space
- Change: capability (New capability)
- Applies to: ERNIE-5.5-Flash-Base

## What changed

Baidu released ERNIE-5.5-Flash-Base, a base model with 280B total and 14B active parameters, with weights in FP8 and BF16.

Baidu has released ERNIE-5.5-Flash-Base, a pure base model with 280B total parameters, 14B of them active for each token, trained on 33.5T tokens. Weights are available in FP8 and BF16. A base model is the pretrained checkpoint before instruction tuning, so it is a foundation for other training rather than a chat model. Baidu's preliminary results against DeepSeek-V4-Flash-Base are 78.4 vs 69.2 on MMLU-Pro, 82.3 vs 69.5 on HumanEval+ and 40.1 vs 31.9 on LiveCodeBench. The release was relayed on X by kimmonismus.

The gap between 280B and 14B comes from the mixture-of-experts design. The network holds many expert sub-networks and routes each token to a few of them, so only a small fraction of the total parameters do the work on any single token. That lowers per-token compute while keeping a large pool of capacity. The post also names ECHO, which runs multi-token prediction through the final layer, but gives no detail on how it is trained.

Attention works differently from most models. Only 6 of the model's 43 layers keep token-level full context. The other layers compress every 128 tokens into a single KV entry. The KV cache stores the key and value vectors that let a model look back at earlier text, and it grows with context length. Keeping full context in only a few layers and compressing the rest is how the design claims up to 1M tokens of context.

The post doesn't say which prompting format, shot count or evaluation harness produced the scores. It compares against one model.

The license reportedly excludes the EU, UK and South Korea.

## Why it matters

Companies in the EU, UK or South Korea would be excluded by the stated license, which matters for any plan to build on these weights.

## What you can do

Confirm the license covers your jurisdiction before evaluating the weights.

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Baidu releases ERNIE-5.5-Flash-Base, a 280B MoE base model trained on 33.5T tokens", 2026-11-07, https://ai-news.ikangai.com/story/baidu-releases-ernie-5-5-flash-base-a-280b-moe-base-model-trained-on-33-5t-token/
